Flowchart of the direction-of-arrival (DOA) estimator. The signals captured by the microphone array are cross-correlated in short temporal segments t using the generalized crosscorrelation (GCC) with phase transform (PHAT). The resulting GCC-PHAT functions p _kl (τ, t) from different microphone pairs (k, l) are combined based on a physically plausible range of delay τ to the feature vector ϕ(t). Each ϕ(t) is classified by a set of R support vector machine (SVM) models trained for different DOA angles α ₁…α _R. The resulting decision values are converted into the source presence probability estimate P(α, t) of direction α via a trained generalized linear model (GLM).